The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/vllm-server.py initializes the vLLM engine with the trust_remote_code=True parameter. This setting enables the execution of arbitrary Python code bundled within a model's repository, posing a significant risk of Remote Code Execution (RCE) if the source model is untrusted or compromised.
[REMOTE_CODE_EXECUTION]: The skill provides patterns for downloading and running LLMs and datasets from external repositories. Specifically, scripts/quantization-config.py fetches datasets from Hugging Face repositories (allenai/c4, codeparrot/github-code) and loads models dynamically, which involves executing remote code components inherent to certain modern model formats.
[COMMAND_EXECUTION]: In scripts/vllm-server.py, the subprocess.Popen function is used to launch an API server. The command is constructed using environment variables such as MODEL_NAME, HOST, and PORT. While the implementation uses a list structure to mitigate basic shell injection, the lack of explicit input validation on these environment variables could be exploited.
[EXTERNAL_DOWNLOADS]: The skill instructions and scripts encourage downloading external resources and tools from the internet. For example, SKILL.md suggests installing portless via npm i -g portless and using npx lighthouse. The LLM inference documentation frequently references downloading model weights from Hugging Face.
[DATA_EXFILTRATION]: The scripts/performance-monitoring.ts file provides a PerformanceReporter class designed to collect and transmit telemetry data to a remote endpoint. While intended for legitimate monitoring, this pattern provides a mechanism that could be misused for data exfiltration if the target URL is maliciously configured.
[INDIRECT_PROMPT_INJECTION]: The skill defines a large attack surface for indirect prompt injection by design. It ingests and processes data from external websites (via agent-browser and WebFetch) and local service logs. A lack of sanitization in the provided templates could allow malicious content in processed data to influence the agent's logic. Evidence found in scripts/performance-monitoring.ts and SKILL.md (Capability inventory: agent-browser, Read, Glob, Grep, WebFetch, WebSearch).

performance